AI Email Follow Up

AI Email Follow Up — independent reviews, comparisons, pricing and step-by-step guides on Aizhi.

  • Semantic compression

    Semantic compression

    In natural language processing, semantic compression is a process of compacting a lexicon used to build a textual document (or a set of documents) by reducing language heterogeneity, while maintaining text semantics. As a result, the same ideas can be represented using a smaller set of words. In most applications, semantic compression is a lossy compression. Increased prolixity does not compensate for the lexical compression and an original document cannot be reconstructed in a reverse process. == By generalization == Semantic compression is basically achieved in two steps, using frequency dictionaries and semantic network: determining cumulated term frequencies to identify target lexicon, replacing less frequent terms with their hypernyms (generalization) from target lexicon. Step 1 requires assembling word frequencies and information on semantic relationships, specifically hyponymy. Moving upwards in word hierarchy, a cumulative concept frequency is calculating by adding a sum of hyponyms' frequencies to frequency of their hypernym: c u m f ( k i ) = f ( k i ) + ∑ j c u m f ( k j ) {\displaystyle cumf(k_{i})=f(k_{i})+\sum _{j}cumf(k_{j})} where k i {\displaystyle k_{i}} is a hypernym of k j {\displaystyle k_{j}} . Then a desired number of words with top cumulated frequencies are chosen to build a target lexicon. In the second step, compression mapping rules are defined for the remaining words in order to handle every occurrence of a less frequent hyponym as its hypernym in output text. Example The below fragment of text has been processed by the semantic compression. Words in bold have been replaced by their hypernyms. They are both nest building social insects, but paper wasps and honey bees organize their colonies in very different ways. In a new study, researchers report that despite their differences, these insects rely on the same network of genes to guide their social behavior.The study appears in the Proceedings of the Royal Society B: Biological Sciences. Honey bees and paper wasps are separated by more than 100 million years of evolution, and there are striking differences in how they divvy up the work of maintaining a colony. The procedure outputs the following text: They are both facility building insect, but insects and honey insects arrange their biological groups in very different structure. In a new study, researchers report that despite their difference of opinions, these insects act the same network of genes to steer their party demeanor. The study appears in the proceeding of the institution bacteria Biological Sciences. Honey insects and insect are separated by more than hundred million years of organic processes, and there are impinging differences of opinions in how they divvy up the work of affirming a biological group. == Implicit semantic compression == A natural tendency to keep natural language expressions concise can be perceived as a form of implicit semantic compression, by omitting unmeaningful words or redundant meaningful words (especially to avoid pleonasms). == Applications and advantages == In the vector space model, compacting a lexicon leads to a reduction of dimensionality, which results in less computational complexity and a positive influence on efficiency. Semantic compression is advantageous in information retrieval tasks, improving their effectiveness (in terms of both precision and recall). This is due to more precise descriptors (reduced effect of language diversity – limited language redundancy, a step towards a controlled dictionary). As in the example above, it is possible to display the output as natural text (re-applying inflexion, adding stop words).

    Read more →
  • E-gree (app)

    E-gree (app)

    E-gree is a legal app that became well known in 2020. It was the first app of its kind to protect users against a number of dating-related issues, including revenge porn. == Background == The app was co-founded by Araz Mamet, Keith Fraser and Ilya Flaks. The app focuses on privacy, with users being able to set up various contracts to protect themselves following a breakup, or while dating. This notably included signing an NDA when sexting. The app received investment from a number of notable people and companies, including Natalia Vodianova.

    Read more →
  • Scale space

    Scale space

    Scale-space theory is a framework for multi-scale signal representation developed by the computer vision, image processing and signal processing communities with complementary motivations from physics and biological vision. It is a formal theory for handling image structures at different scales, by representing an image as a one-parameter family of smoothed images, the scale-space representation, parametrized by the size of the smoothing kernel used for suppressing fine-scale structures. The parameter t {\displaystyle t} in this family is referred to as the scale parameter, with the interpretation that image structures of spatial size smaller than about t {\displaystyle {\sqrt {t}}} have largely been smoothed away in the scale-space level at scale t {\displaystyle t} . The main type of scale space is the linear (Gaussian) scale space, which has wide applicability as well as the attractive property of being possible to derive from a small set of scale-space axioms. The corresponding scale-space framework encompasses a theory for Gaussian derivative operators, which can be used as a basis for expressing a large class of visual operations for computerized systems that process visual information. This framework also allows visual operations to be made scale invariant, which is necessary for dealing with the size variations that may occur in image data, because real-world objects may be of different sizes and in addition the distance between the object and the camera may be unknown and may vary depending on the circumstances. == Definition == The notion of scale space applies to signals of arbitrary numbers of variables. The most common case in the literature applies to two-dimensional images, which is what is presented here. Consider a given image f {\displaystyle f} where f ( x , y ) {\displaystyle f(x,y)} is the greyscale value of the pixel at position ( x , y ) {\displaystyle (x,y)} . The linear (Gaussian) scale-space representation of f {\displaystyle f} is a family of derived signals L ( x , y ; t ) {\displaystyle L(x,y;t)} defined by the convolution of f ( x , y ) {\displaystyle f(x,y)} with the two-dimensional Gaussian kernel g ( x , y ; t ) = 1 2 π t e − ( x 2 + y 2 ) / 2 t {\displaystyle g(x,y;t)={\frac {1}{2\pi t}}e^{-(x^{2}+y^{2})/2t}\,} such that L ( ⋅ , ⋅ ; t ) = g ( ⋅ , ⋅ ; t ) ∗ f ( ⋅ , ⋅ ) , {\displaystyle L(\cdot ,\cdot ;t)\ =g(\cdot ,\cdot ;t)f(\cdot ,\cdot ),} where the semicolon in the argument of L {\displaystyle L} implies that the convolution is performed only over the variables x , y {\displaystyle x,y} , while the scale parameter t {\displaystyle t} after the semicolon just indicates which scale level is being defined. This definition of L {\displaystyle L} works for a continuum of scales t ≥ 0 {\displaystyle t\geq 0} , but typically only a finite discrete set of levels in the scale-space representation would be actually considered. The scale parameter t = σ 2 {\displaystyle t=\sigma ^{2}} is the variance of the Gaussian filter and as a limit for t = 0 {\displaystyle t=0} the filter g {\displaystyle g} becomes an impulse function such that L ( x , y ; 0 ) = f ( x , y ) , {\displaystyle L(x,y;0)=f(x,y),} that is, the scale-space representation at scale level t = 0 {\displaystyle t=0} is the image f {\displaystyle f} itself. As t {\displaystyle t} increases, L {\displaystyle L} is the result of smoothing f {\displaystyle f} with a larger and larger filter, thereby removing more and more of the details that the image contains. Since the standard deviation of the filter is σ = t {\displaystyle \sigma ={\sqrt {t}}} , details that are significantly smaller than this value are to a large extent removed from the image at scale parameter t {\displaystyle t} , see the following figures and for graphical illustrations. === Why a Gaussian filter? === When faced with the task of generating a multi-scale representation one may ask: could any filter g of low-pass type and with a parameter t which determines its width be used to generate a scale space? The answer is no, as it is of crucial importance that the smoothing filter does not introduce new spurious structures at coarse scales that do not correspond to simplifications of corresponding structures at finer scales. In the scale-space literature, a number of different ways have been expressed to formulate this criterion in precise mathematical terms. The conclusion from several different axiomatic derivations that have been presented is that the Gaussian scale space constitutes the canonical way to generate a linear scale space, based on the essential requirement that new structures must not be created when going from a fine scale to any coarser scale. Conditions, referred to as scale-space axioms, that have been used for deriving the uniqueness of the Gaussian kernel include linearity, shift invariance, semi-group structure, non-enhancement of local extrema, scale invariance and rotational invariance. In the works, the uniqueness claimed in the arguments based on scale invariance has been criticized, and alternative self-similar scale-space kernels have been proposed. The Gaussian kernel is, however, a unique choice according to the scale-space axiomatics based on causality or non-enhancement of local extrema. === Alternative definition === Equivalently, the scale-space family can be defined as the solution of the diffusion equation (for example in terms of the heat equation), ∂ t L = 1 2 ∇ 2 L , {\displaystyle \partial _{t}L={\frac {1}{2}}\nabla ^{2}L,} with initial condition L ( x , y ; 0 ) = f ( x , y ) {\displaystyle L(x,y;0)=f(x,y)} . This formulation of the scale-space representation L means that it is possible to interpret the intensity values of the image f as a "temperature distribution" in the image plane and that the process that generates the scale-space representation as a function of t corresponds to heat diffusion in the image plane over time t (assuming the thermal conductivity of the material equal to the arbitrarily chosen constant ⁠1/2⁠). Although this connection may appear superficial for a reader not familiar with differential equations, it is indeed the case that the main scale-space formulation in terms of non-enhancement of local extrema is expressed in terms of a sign condition on partial derivatives in the 2+1-D volume generated by the scale space, thus within the framework of partial differential equations. Furthermore, a detailed analysis of the discrete case shows that the diffusion equation provides a unifying link between continuous and discrete scale spaces, which also generalizes to nonlinear scale spaces, for example, using anisotropic diffusion. Hence, one may say that the primary way to generate a scale space is by the diffusion equation, and that the Gaussian kernel arises as the Green's function of this specific partial differential equation. == Motivations == The motivation for generating a scale-space representation of a given data set originates from the basic observation that real-world objects are composed of different structures at different scales. This implies that real-world objects, in contrast to idealized mathematical entities such as points or lines, may appear in different ways depending on the scale of observation. For example, the concept of a "tree" is appropriate at the scale of meters, while concepts such as leaves and molecules are more appropriate at finer scales. For a computer vision system analysing an unknown scene, there is no way to know a priori what scales are appropriate for describing the interesting structures in the image data. Hence, the only reasonable approach is to consider descriptions at multiple scales in order to be able to capture the unknown scale variations that may occur. Taken to the limit, a scale-space representation considers representations at all scales. Another motivation to the scale-space concept originates from the process of performing a physical measurement on real-world data. In order to extract any information from a measurement process, one has to apply operators of non-infinitesimal size to the data. In many branches of computer science and applied mathematics, the size of the measurement operator is disregarded in the theoretical modelling of a problem. The scale-space theory on the other hand explicitly incorporates the need for a non-infinitesimal size of the image operators as an integral part of any measurement as well as any other operation that depends on a real-world measurement. There is a close link between scale-space theory and biological vision. Many scale-space operations show a high degree of similarity with receptive field profiles recorded from the mammalian retina and the first stages in the visual cortex. In these respects, the scale-space framework can be seen as a theoretically well-founded paradigm for early vision, which in addition has been thoroughly tested by algorithms and experiments. == Gaussian derivatives == At any scale in scale space, we c

    Read more →
  • Synthesia (company)

    Synthesia (company)

    Synthesia Limited is a British multinational artificial intelligence company based in London, United Kingdom. It is a synthetic media-generation software developer and creator of AI-generated video content, including audio-visual agents and cloned avatars. Britain's largest generative-AI firm, it is used by 70% of FTSE 100 and over 90% of Fortune 100 companies. == Overview == Synthesia is most often used by corporations for localized communication, orientation, employee training videos, advertising campaigns, reporting, product demonstrations, customer service, and to create chatbots. Its software algorithm mimics speech and facial movements based on video recordings of an individual’s speech and facial expressions. From this, a text-to-speech video is created to look and sound like the individual. Swiss bank UBS incorporated Synthesia AI-powered avatars of their human financial experts, for instance, in 2025. Users create content via the platform's pre-generated AI presenters or by creating digital representations of themselves, or personal avatars, using the platform's AI video editing tool. These avatars can be used to narrate videos generated from text. As of August 2021, Synthesia's voice database included multiple gender options in over 60 languages. Its free voice library doubled by 2025, to 140 languages and accents, and its Express-Voice technology can clone a user's own voice, or generate a synthetic one. === Deepfakes === The platform prohibits use of its software to create non-consensual clones, including of celebrities or political figures for satirical purposes. Explicit consent must be provided in addition to a strict pre-screening regimen for use of an individual's likeness to avoid “deepfaking”. While the company prohibits use of its technology for misinformation or "news-like content", an October 2023 Freedom House report stated that Synthesia tools had been used by governments in Venezuela, China, Burkina Faso, and Russia to create videos of fake TV news outlets with AI-generated avatars in order to spread propaganda. Actor Dan Dewhirst signed a contract with the company in 2021, becoming one of the first actors whose likeness would be made into an AI avatar, finding his likeness used in the Venezuelan generated-videos. The company stated, in February 2024, that it had improved its misuse detection systems, and, in April 2024, that new users of its technology are screened by the company, and content employing it is further vetted by Synthesia moderators. == History == Synthesia's software utilizes deep learning architecture developed by Lourdes Agapito and Matthias Niessner. The company was co-founded in 2017 by Agapito, Niessner, Victor Riparbelli, and Steffen Tjerrild. In 2018, the company first demonstrated the software's capabilities on the BBC programme Click when it presented a digitization of Matthew Amroliwala speaking Spanish, Mandarin, and Hindi. Through Synthesia's first two years of existence, it employed 10 people and struggled to make sales, leading to an expansion of the company's focus. It moved on from just targeting entertainment studios to a variety of businesses. In 2020, Synthesia users were reported to include Amazon, Tiffany & Co. and IHG Hotels & Resorts. In January 2024, the company introduced its AI video assistant, which turns text-to-video. That April, with a reported 55,000 customers, including half of the Fortune 100, Synthesia launched "expressive avatars". That September, an enhanced dubbing feature was launched, to translate video in 30 languages with naturalized lip-syncing. Peter Hill joined Synthesia as CTO in January 2025, following 25 years at Amazon, and two years as CEO and CPO of Wildfire Studios. That March, a million dollar base of shares was formed to furnish human actors, employed to generate digital avatars, with company stock, which all of its employees hold. By June of that year, 150,000 individuals from among Synthesia's 65,000 customers had created AI-generated avatars of themselves. In July 2025, the company's new global headquarters at Regent’s Place was opened by London mayor Sadiq Khan, who described Britain's largest generative-AI company, then valued at over $2 billion, as a "London success story". By that October, its technology was employed by 90% of the Fortune 100, and Synthesia 3.0 was launched, with hyper-realistic digital avatars equipped with AI-powered dubbing and translation, and a built-in video assistant. In January 2026, it reached a $4 billion valuation, with 70% of FTSE 100 companies noted among its customers. === Funding === The company raised $3.1 million in seed funding in 2019. In April 2021, the company raised $12.5 million in Series A funding. In December 2021, it raised $50 million in a Series B funding round led by Kleiner Perkins and GV (then Google Ventures). Synthesia gained a total valuation of $1 billion, and achieved unicorn status, when it raised $90 million from Accel and Nvidia partnership NVentures, in June 2023, during its Series C funding round. Counting 60,000 customers by January 2025, including over 60% of Fortune 100 companies; the company raised $180 million in a Series D round led by NEA, with new investors World Innovation Lab (WiL), Atlassian Ventures and PSP Growth, as well as existing investors GV, MMC Ventures and FirstMark, doubling Synthesia's valuation to $2.1 billion. Capital raised by 2025 had reached $330 million, with investments slated to further product innovation, talent growth, and company expansion in North America, Europe, Japan and Australia. In April 2025, Adobe Inc. invested £10 million in the company for a strategic partnership. Synthesia subsequently rejected a $3 billion acquisition offer from Adobe, choosing to remain independent. With a revenue stream then exceeding $100 million annually; GV led a Series E funding round in October 2025, resulting in Synthesia's $4 billion valuation, raising $200 million from GV, Nvidia and Accel to develop, in 2026, interactive audio-visual avatar "agents" that converse on topic, for automated sales training and corporate communications, such as recruiting. == Recognition == In 2021, Synthesia partnered with Lay's to create the Messi Messages campaign featuring Argentine footballer Lionel Messi. Users created personalized messages with Synthesia's software and sent custom artificial reality video messages from Messi based on their text input. The campaign received a Cannes Lion Award under the Bronze category. In February 2025, UK Science and Technology Minister Peter Kyle commended Synthesia's "pioneering generative AI innovations."

    Read more →
  • MobileNet

    MobileNet

    MobileNet is a family of convolutional neural network (CNN) architectures designed for image classification, object detection, and other computer vision tasks. They are designed for small size, low latency, and low power consumption, making them suitable for on-device inference and edge computing on resource-constrained devices like mobile phones and embedded systems. They were originally designed to be run efficiently on mobile devices with TensorFlow Lite. The need for efficient deep learning models on mobile devices led researchers at Google to develop MobileNet. As of June 2025, the family has five versions, each improving upon the previous one in terms of performance and efficiency. == Features == === V1 === MobileNetV1 was published in April 2017. Its main architectural innovation was incorporation of depthwise separable convolutions. It was first developed by Laurent Sifre during an internship at Google Brain in 2013 as an architectural variation on AlexNet to improve convergence speed and model size. The depthwise separable convolution decomposes a single standard convolution into two convolutions: a depthwise convolution that filters each input channel independently and a pointwise convolution ( 1 × 1 {\displaystyle 1\times 1} convolution) that combines the outputs of the depthwise convolution. This factorization significantly reduces computational cost. The MobileNetV1 has two hyperparameters: a width multiplier α {\displaystyle \alpha } that controls the number of channels in each layer. Smaller values of α {\displaystyle \alpha } lead to smaller and faster models, but at the cost of reduced accuracy, and a resolution multiplier ρ {\displaystyle \rho } , which controls the input resolution of the images. Lower resolutions result in faster processing but potentially lower accuracy. === V2 === MobileNetV2 was published in March 2019. It uses inverted residual layers and linear bottlenecks. Inverted residuals modify the traditional residual block structure. Instead of compressing the input channels before the depthwise convolution, they expand them. This expansion is followed by a 1 × 1 {\displaystyle 1\times 1} depthwise convolution and then a 1 × 1 {\displaystyle 1\times 1} projection layer that reduces the number of channels back down. This inverted structure helps to maintain representational capacity by allowing the depthwise convolution to operate on a higher-dimensional feature space, thus preserving more information flow during the convolutional process. Linear bottlenecks removes the typical ReLU activation function in the projection layers. This was rationalized by arguing that that nonlinear activation loses information in lower-dimensional spaces, which is problematic when the number of channels is already small. === V3 === MobileNetV3 was published in 2019. The publication included MobileNetV3-Small, MobileNetV3-Large, and MobileNetEdgeTPU (optimized for Pixel 4). They were found by a form of neural architecture search (NAS) that takes mobile latency into account, to achieve good trade-off between accuracy and latency. It used piecewise-linear approximations of swish and sigmoid activation functions (which they called "h-swish" and "h-sigmoid"), squeeze-and-excitation modules, and the inverted bottlenecks of MobileNetV2. === V4 === MobileNetV4 was published in September 2024. The publication included a large number of architectures found by NAS. Inspired by Vision Transformers, the V4 series included multi-query attention. It also unified both inverted residual and inverted bottleneck from the V3 series with the "universal inverted bottleneck", which includes these two as special cases. === V5 === MobileNetV5's architecture was published shortly after the release of Gemma 3n in June 2025. While the announcement stated a technical report on MobileNetV5 would be available soon, this has not yet materialised. The network is 10 times larger than the largest V4 variant.

    Read more →
  • Visual temporal attention

    Visual temporal attention

    Visual temporal attention is a special case of visual attention that involves directing attention to specific instant of time. Similar to its spatial counterpart visual spatial attention, these attention modules have been widely implemented in video analytics in computer vision to provide enhanced performance and human interpretable explanation of deep learning models. As visual spatial attention mechanism allows human and/or computer vision systems to focus more on semantically more substantial regions in space, visual temporal attention modules enable machine learning algorithms to emphasize more on critical video frames in video analytics tasks, such as human action recognition. In convolutional neural network-based systems, the prioritization introduced by the attention mechanism is regularly implemented as a linear weighting layer with parameters determined by labeled training data. == Application in Action Recognition == Recent video segmentation algorithms often exploits both spatial and temporal attention mechanisms. Research in human action recognition has accelerated significantly since the introduction of powerful tools such as Convolutional Neural Networks (CNNs). However, effective methods for incorporation of temporal information into CNNs are still being actively explored. Motivated by the popular recurrent attention models in natural language processing, the Attention-aware Temporal Weighted CNN (ATW CNN) is proposed in videos, which embeds a visual attention model into a temporal weighted multi-stream CNN. This attention model is implemented as temporal weighting and it effectively boosts the recognition performance of video representations. Besides, each stream in the proposed ATW CNN framework is capable of end-to-end training, with both network parameters and temporal weights optimized by stochastic gradient descent (SGD) with back-propagation. Experimental results show that the ATW CNN attention mechanism contributes substantially to the performance gains with the more discriminative snippets by focusing on more relevant video segments. == Literature == Seibold VC, Balke J and Rolke B (2023): Temporal attention. Front. Cognit. 2:1168320. doi: 10.3389/fcogn.2023.1168320.

    Read more →
  • Digital image processing

    Digital image processing

    Digital image processing is the use of a digital computer to process digital images through an algorithm. As a subcategory or field of digital signal processing, digital image processing has many advantages over analog image processing. It allows a much wider range of algorithms to be applied to the input data and can avoid problems such as the build-up of noise and distortion during processing. Since images are defined over two dimensions (perhaps more), digital image processing may be modeled in the form of multidimensional systems. The generation and development of digital image processing are mainly affected by three factors: first, the development of computers; second, the development of mathematics (especially the creation and improvement of discrete mathematics theory); and third, the demand for a wide range of applications in environment, agriculture, military, industry and medical science has increased. == History == Many of the techniques of digital image processing, or digital picture processing as it often was called, were developed in the 1960s, at Bell Laboratories, the Jet Propulsion Laboratory, Massachusetts Institute of Technology, University of Maryland, and a few other research facilities, with application to satellite imagery, wire-photo standards conversion, medical imaging, videophone, character recognition, and photograph enhancement. The purpose of early image processing was to improve the quality of the image. In image processing, the input is a low-quality image, and the output is an image with improved quality. Common image processing includes image enhancement, restoration, encoding, and compression. The first successful application was the American Jet Propulsion Laboratory (JPL). They used image processing techniques such as geometric correction, gradation transformation, noise removal, etc. on the thousands of lunar photos sent back by the Space Detector Ranger 7 in 1964, taking into account the position of the Sun and the environment of the Moon. The impact of the successful mapping of the Moon's surface map by the computer has been a success. Later, more complex image processing was performed on the nearly 100,000 photos sent back by the spacecraft, so that the topographic map, color map and panoramic mosaic of the Moon were obtained, which achieved extraordinary results and laid a solid foundation for human landing on the Moon. The cost of processing was fairly high, however, with the computing equipment of that era. That changed in the 1970s, when digital image processing proliferated as cheaper computers and dedicated hardware became available. This led to images being processed in real-time, for some dedicated problems such as television standards conversion. As general-purpose computers became faster, they started to take over the role of dedicated hardware for all but the most specialized and computer-intensive operations. With the fast computers and signal processors available in the 2000s, digital image processing has become the most common form of image processing, and is generally used because it is not only the most versatile method, but also the cheapest. === Image sensors === The basis for modern image sensors is metal–oxide–semiconductor (MOS) technology, invented at Bell Labs between 1955 and 1960, This led to the development of digital semiconductor image sensors, including the charge-coupled device (CCD) and later the CMOS sensor. The charge-coupled device was invented by Willard S. Boyle and George E. Smith at Bell Labs in 1969. While researching MOS technology, they realized that an electric charge was the analogy of the magnetic bubble and that it could be stored on a tiny MOS capacitor. As it was fairly straightforward to fabricate a series of MOS capacitors in a row, they connected a suitable voltage to them so that the charge could be stepped along from one to the next. The CCD is a semiconductor circuit that was later used in the first digital video cameras for television broadcasting. The NMOS active-pixel sensor (APS) was invented by Olympus in Japan during the mid-1980s. This was enabled by advances in MOS semiconductor device fabrication, with MOSFET scaling reaching smaller micron and then sub-micron levels. The NMOS APS was fabricated by Tsutomu Nakamura's team at Olympus in 1985. The CMOS active-pixel sensor (CMOS sensor) was later developed by Eric Fossum's team at the NASA Jet Propulsion Laboratory in 1993. By 2007, sales of CMOS sensors had surpassed CCD sensors. MOS image sensors are widely used in optical mouse technology. The first optical mouse, invented by Richard F. Lyon at Xerox in 1980, used a 5 μm NMOS integrated circuit sensor chip. Since the first commercial optical mouse, the IntelliMouse introduced in 1999, most optical mouse devices use CMOS sensors. === Image compression === An important development in digital image compression technology was the discrete cosine transform (DCT), a lossy compression technique first proposed by Nasir Ahmed in 1972. DCT compression became the basis for JPEG, which was introduced by the Joint Photographic Experts Group in 1992. JPEG compresses images down to much smaller file sizes, and has become the most widely used image file format on the Internet. Its highly efficient DCT compression algorithm was largely responsible for the wide proliferation of digital images and digital photos, with several billion JPEG images produced every day as of 2015. Medical imaging techniques produce very large amounts of data, especially from CT, MRI and PET modalities. As a result, storage and communications of electronic image data are prohibitive without the use of compression. JPEG 2000 image compression is used by the DICOM standard for storage and transmission of medical images. The cost and feasibility of accessing large image data sets over low or various bandwidths are further addressed by use of another DICOM standard, called JPIP, to enable efficient streaming of the JPEG 2000 compressed image data. === Digital signal processor (DSP) === Electronic signal processing was revolutionized by the wide adoption of MOS technology in the 1970s. MOS integrated circuit technology was the basis for the first single-chip microprocessors and microcontrollers in the early 1970s, and then the first single-chip digital signal processor (DSP) chips in the late 1970s. DSP chips have since been widely used in digital image processing. The discrete cosine transform (DCT) image compression algorithm has been widely implemented in DSP chips, with many companies developing DSP chips based on DCT technology. DCTs are widely used for encoding, decoding, video coding, audio coding, multiplexing, control signals, signaling, analog-to-digital conversion, formatting luminance and color differences, and color formats such as YUV444 and YUV411. DCTs are also used for encoding operations such as motion estimation, motion compensation, inter-frame prediction, quantization, perceptual weighting, entropy encoding, variable encoding, and motion vectors, and decoding operations such as the inverse operation between different color formats (YIQ, YUV and RGB) for display purposes. DCTs are also commonly used for high-definition television (HDTV) encoder/decoder chips. == Tasks == Digital image processing allows the use of much more complex algorithms, and hence, can offer both more sophisticated performance at simple tasks, and the implementation of methods which would be impossible by analogue means. In particular, digital image processing is a concrete application of, and a practical technology based on: Classification Feature extraction Multi-scale signal analysis Pattern recognition Projection Some techniques that are used in digital image processing include: Anisotropic diffusion Hidden Markov models Image editing Image restoration Independent component analysis Linear filtering Neural networks Partial differential equations Pixelation Point feature matching Principal components analysis Self-organizing maps Wavelets == Digital image transformations == === Filtering === Digital filters are used to blur and sharpen digital images. Filtering can be performed by: convolution with specifically designed kernels (filter array) in the spatial domain masking specific frequency regions in the frequency (Fourier) domain The following examples show both methods: ==== Image padding in Fourier domain filtering ==== Images are typically padded before being transformed to the Fourier space, the highpass filtered images below illustrate the consequences of different padding techniques: Notice that the highpass filter shows extra edges when zero padded compared to the repeated edge padding. ==== Filtering code examples ==== MATLAB example for spatial domain highpass filtering. === Affine transformations === Affine transformations enable basic image transformations including scale, rotate, translate, mirror and shear as is shown in the following examples: To apply the affine

    Read more →
  • PhyCV

    PhyCV

    PhyCV is the first computer vision library which utilizes algorithms directly derived from the equations of physics governing physical phenomena. The algorithms appearing in the first release emulate the propagation of light through a physical medium with natural and engineered diffractive properties followed by coherent detection. Unlike traditional algorithms that are a sequence of hand-crafted empirical rules, physics-inspired algorithms leverage physical laws of nature as blueprints. In addition, these algorithms can, in principle, be implemented in real physical devices for fast and efficient computation in the form of analog computing. Currently PhyCV has three algorithms, Phase-Stretch Transform (PST) and Phase-Stretch Adaptive Gradient-Field Extractor (PAGE), and Vision Enhancement via Virtual diffraction and coherent Detection (VEViD). All algorithms have CPU and GPU versions. PhyCV is now available on GitHub and can be installed from pip. == History == Algorithms in PhyCV are inspired by the physics of the photonic time stretch (a hardware technique for ultrafast and single-shot data acquisition). PST is an edge detection algorithm that was open-sourced in 2016 and has 800+ stars and 200+ forks on GitHub. PAGE is a directional edge detection algorithm that was open-sourced in February, 2022. PhyCV was originally developed and open-sourced by Jalali-Lab @ UCLA in May 2022. In the initial release of PhyCV, the original open-sourced code of PST and PAGE is significantly refactored and improved to be modular, more efficient, GPU-accelerated and object-oriented. VEViD is a low-light and color enhancement algorithm that was added to PhyCV in November 2022. == Background == === Phase-Stretch Transform (PST) === Phase-Stretch Transform (PST) is a computationally efficient edge and texture detection algorithm with exceptional performance in visually impaired images. The algorithm transforms the image by emulating propagation of light through a device with engineered diffractive property followed by coherent detection. It has been applied in improving the resolution of MRI image, extracting blood vessels in retina images, dolphin identification, and waste water treatment, single molecule biological imaging, and classification of UAV using micro Doppler imaging. === Phase-Stretch Adaptive Gradient-Field Extractor (PAGE) === Phase-Stretch Adaptive Gradient-Field Extractor (PAGE) is a physics-inspired algorithm for detecting edges and their orientations in digital images at various scales. The algorithm is based on the diffraction equations of optics. Metaphorically speaking, PAGE emulates the physics of birefringent (orientation-dependent) diffractive propagation through a physical device with a specific diffractive structure. The propagation converts a real-valued image into a complex function. Related information is contained in the real and imaginary components of the output. The output represents the phase of the complex function. === Vision Enhancement via Virtual diffraction and coherent Detection (VEViD) === Vision Enhancement via Virtual diffraction and coherent Detection (VEViD) an efficient and interpretable low-light and color enhancement algorithm that reimagines a digital image as a spatially varying metaphoric light field and then subjects the field to the physical processes akin to diffraction and coherent detection. The term “Virtual” captures the deviation from the physical world. The light field is pixelated and the propagation imparts a phase with an arbitrary dependence on frequency which can be different from the quadratic behavior of physical diffraction. VEViD can be further accelerated through mathematical approximations that reduce the computation time without appreciable sacrifice in image quality. A closed-form approximation for VEViD which we call VEViD-lite can achieve up to 200 FPS for 4K video enhancement. == PhyCV on the Edge == Featuring low-dimensionality and high-efficiency, PhyCV is ideal for edge computing applications. In this section, we demonstrate running PhyCV on NVIDIA Jetson Nano in real-time. === NVIDIA Jetson Nano Developer Kit === NVIDIA Jetson Nano Developer Kit is a small- sized and power-efficient platform for edge computing applications. It is equipped with an NVIDIA Maxwell architecture GPU with 128 CUDA cores, a quad-core ARM Cortex-A57 CPU, 4GB 64-bit LPDDR4 RAM, and supports video encoding and decoding up to 4K resolution. Jetson Nano also offers a variety of interfaces for connectivity and expansion, making it ideal for a wide range of AI and IoT applications. In our setup, we connect a USB camera to the Jetson Nano to acquire videos and demonstrate using PhyCV to process the videos in real-time. === Real-time PhyCV on Jetson Nano === We use the Jetson Nano (4GB) with NVIDIA JetPack SDK version 4.6.1, which comes with pre- installed Python 3.6, CUDA 10.2, and OpenCV 4.1.1. We further install PyTorch 1.10 to enable the GPU accelerated PhyCV. We demonstrate the results and metrics of running PhyCV on Jetson Nano in real-time for edge detection and low-light enhancement tasks. For 480p videos, both operations achieve beyond 38 FPS, which is sufficient for most cameras that capture videos at 30 FPS. For 720p videos, PhyCV low-light enhancement can operate at 24 FPS and PhyCV edge detection can operate at 17 FPS. == Highlights == === Modular Code Architecture === The code in PhyCV has a modular design which faithfully follows the physical process from which the algorithm was originated. Both PST and PAGE modules in the PhyCV library emulate the propagation of the input signal (original digital image) through a device with engineered diffractive property followed by coherent (phase) detection. The dispersive propagation applies a phase kernel to the frequency domain of the original image. This process has three steps in general, loading the image, initializing the kernel and applying the kernel. In the implementation of PhyCV, each algorithm is represented as a class in Python and each class has methods that simulate the steps described above. The modular code architecture follows the physics behind the algorithm. Please refer to the source code on GitHub for more details. === GPU Acceleration === PhyCV supports GPU acceleration. The GPU versions of PST and PAGE are built on PyTorch accelerated by the CUDA toolkit. The acceleration is beneficial for applying the algorithms in real-time image video processing and other deep learning tasks. The running time per frame of PhyCV algorithms on CPU (Intel i9-9900K) and GPU (NVIDIA TITAN RTX) for videos at different resolutions are shown below. Note that the PhyCV low-light enhancement operates in the HSV color space, so the running time also includes RGB to HSV conversion. However, for all running times using GPUs, we ignore the time of moving data from CPUs to GPUs and count the algorithm operation time only. == Installation and Examples == Please refer to the GitHub README file for a detailed technical documentation. == Current Limitations == === I/O (Input/Output) Bottleneck for Real-time Video Processing === When dealing with real-time video streams from cameras, the frames are captured and buffered in CPU and have to be moved to GPU to run the GPU-accelerated PhyCV algorithms. This process is time-consuming and it is a common bottleneck for real-time video-processing algorithms. === Lack of Parameter Adaptivity for Different Images === Currently, the parameters of PhyCV algorithms have to be manually tuned for different images. Although a set of pre-selected parameters work relatively well for a wide range of images, the lack of parameter adaptivity for different images remains a limitation for now.

    Read more →
  • Poop Map

    Poop Map

    Poop Map is a social app where users can track on a map where and when they defecate. In addition to logging location and time of each bowel movement, users can also add a photo, "like" other users' logs, and rate each account. The social elements of the app allow for groups of users to create a competitive league. Certain behaviors unlock achievements in-app. == Development == The app was created by app developer Nino Uzelac. It was launched in July 2013. == Popularity == The app charted at number one on the Apple App Store charts in 2021 after going viral on TikTok. As of September 2024, the app has a 4.8 rating on the App Store and more than 58,000 ratings. It also has more than one million downloads on the Google Play Store. Poop Map is notably popular among hikers, and has been written about in the outdoors magazine Outside.

    Read more →
  • Rule-based machine translation

    Rule-based machine translation

    Rule-based machine translation (RBMT) is a classical approach of machine translation systems based on linguistic information about source and target languages. Such information is retrieved from (unilingual, bilingual or multilingual) dictionaries and grammars covering the main semantic, morphological, and syntactic regularities of each language. Having input sentences, an RBMT system generates output sentences on the basis of analysis of both the source and the target languages involved. RBMT has been progressively superseded by more efficient methods, particularly neural machine translation. == History == The first RBMT systems were developed in the early 1970s. The most important steps of this evolution were the emergence of the following RBMT systems: Systran Japanese MT systems Today, other common RBMT systems include: Apertium GramTrans == Types of RBMT == There are three different types of rule-based machine translation systems: Direct Systems (Dictionary Based Machine Translation) map input to output with basic rules. Transfer RBMT Systems (Transfer Based Machine Translation) employ morphological and syntactical analysis. Interlingual RBMT Systems (Interlingua) use an abstract meaning. RBMT systems can also be characterized as the systems opposite to Example-based Systems of Machine Translation (Example Based Machine Translation), whereas Hybrid Machine Translations Systems make use of many principles derived from RBMT. == Basic principles == The main approach of RBMT systems is based on linking the structure of the given input sentence with the structure of the demanded output sentence, necessarily preserving their unique meaning. The following example can illustrate the general frame of RBMT: A girl eats an apple. Source Language = English; Demanded Target Language = German Minimally, to get a German translation of this English sentence one needs: A dictionary that will map each English word to an appropriate German word. Rules representing regular English sentence structure. Rules representing regular German sentence structure. And finally, we need rules according to which one can relate these two structures together. Accordingly, we can state the following stages of translation: 1st: getting basic part-of-speech information of each source word: a = indef.article; girl = noun; eats = verb; an = indef.article; apple = noun 2nd: getting syntactic information about the verb "to eat": NP-eat-NP; here: eat – Present Simple, 3rd Person Singular, Active Voice 3rd: parsing the source sentence: (NP an apple) = the object of eat Often only partial parsing is sufficient to get to the syntactic structure of the source sentence and to map it onto the structure of the target sentence. 4th: translate English words into German a (category = indef.article) => ein (category = indef.article) girl (category = noun) => Mädchen (category = noun) eat (category = verb) => essen (category = verb) an (category = indef. article) => ein (category = indef.article) apple (category = noun) => Apfel (category = noun) 5th: Mapping dictionary entries into appropriate inflected forms (final generation): A girl eats an apple. => Ein Mädchen isst einen Apfel. == Ontologies == An ontology is a formal representation of knowledge that includes the concepts (such as objects, processes etc.) in a domain and some relations between them. If the stored information is of linguistic nature, one can speak of a lexicon. In NLP, ontologies can be used as a source of knowledge for machine translation systems. With access to a large knowledge base, rule-based systems can be enabled to resolve many (especially lexical) ambiguities on their own. In the following classic examples, as humans, we are able to interpret the prepositional phrase according to the context because we use our world knowledge, stored in our lexicons:I saw a man/star/molecule with a microscope/telescope/binoculars.Since the syntax does not change, a traditional rule-based machine translation system may not be able to differentiate between the meanings. With a large enough ontology as a source of knowledge however, the possible interpretations of ambiguous words in a specific context can be reduced. === Building ontologies === The ontology generated for the PANGLOSS knowledge-based machine translation system in 1993 may serve as an example of how an ontology for NLP purposes can be compiled: A large-scale ontology is necessary to help parsing in the active modules of the machine translation system. In the PANGLOSS example, about 50,000 nodes were intended to be subsumed under the smaller, manually-built upper (abstract) region of the ontology. Because of its size, it had to be created automatically. The goal was to merge the two resources LDOCE online and WordNet to combine the benefits of both: concise definitions from Longman, and semantic relations allowing for semi-automatic taxonomization to the ontology from WordNet. A definition match algorithm was created to automatically merge the correct meanings of ambiguous words between the two online resources, based on the words that the definitions of those meanings have in common in LDOCE and WordNet. Using a similarity matrix, the algorithm delivered matches between meanings including a confidence factor. This algorithm alone, however, did not match all meanings correctly on its own. A second hierarchy match algorithm was therefore created which uses the taxonomic hierarchies found in WordNet (deep hierarchies) and partially in LDOCE (flat hierarchies). This works by first matching unambiguous meanings, then limiting the search space to only the respective ancestors and descendants of those matched meanings. Thus, the algorithm matched locally unambiguous meanings (for instance, while the word seal as such is ambiguous, there is only one meaning of seal in the animal subhierarchy). Both algorithms complemented each other and helped constructing a large-scale ontology for the machine translation system. The WordNet hierarchies, coupled with the matching definitions of LDOCE, were subordinated to the ontology's upper region. As a result, the PANGLOSS MT system was able to make use of this knowledge base, mainly in its generation element. == Components == The RBMT system contains: a SL morphological analyser - analyses a source language word and provides the morphological information; a SL parser - is a syntax analyser which analyses source language sentences; a translator - used to translate a source language word into the target language; a TL morphological generator - works as a generator of appropriate target language words for the given grammatica information; a TL parser - works as a composer of suitable target language sentences; Several dictionaries - more specifically a minimum of three dictionaries: a SL dictionary - needed by the source language morphological analyser for morphological analysis, a bilingual dictionary - used by the translator to translate source language words into target language words, a TL dictionary - needed by the target language morphological generator to generate target language words. The RBMT system makes use of the following: a Source Grammar for the input language which builds syntactic constructions from input sentences; a Source Lexicon which captures all of the allowable vocabulary in the domain; Source Mapping Rules which indicate how syntactic heads and grammatical functions in the source language are mapped onto domain concepts and semantic roles in the interlingua; a Domain Model/Ontology which defines the classes of domain concepts and restricts the fillers of semantic roles for each class; Target Mapping Rules which indicate how domain concepts and semantic roles in the interlingua are mapped onto syntactic heads and grammatical functions in the target language; a Target Lexicon which contains appropriate target lexemes for each domain concept; a Target Grammar for the target language which realizes target syntactic constructions as linearized output sentences. == Advantages == No bilingual texts are required. This makes it possible to create translation systems for languages that have no texts in common, or even no digitized data whatsoever. Domain independent. Rules are usually written in a domain independent manner, so the vast majority of rules will "just work" in every domain, and only a few specific cases per domain may need rules written for them. No quality ceiling. Every error can be corrected with a targeted rule, even if the trigger case is extremely rare. This is in contrast to statistical systems where infrequent forms will be washed away by default. Total control. Because all rules are hand-written, you can easily debug a rule-based system to see exactly where a given error enters the system, and why. Reusability. Because RBMT systems are generally built from a strong source language analysis that is fed to a transfer step and target language generator, the source language analysis and targe

    Read more →
  • Scale-space axioms

    Scale-space axioms

    In image processing and computer vision, a scale space framework can be used to represent an image as a family of gradually smoothed images. This framework is very general and a variety of scale space representations exist. A typical approach for choosing a particular type of scale space representation is to establish a set of scale-space axioms, describing basic properties of the desired scale-space representation and often chosen so as to make the representation useful in practical applications. Once established, the axioms narrow the possible scale-space representations to a smaller class, typically with only a few free parameters. A set of standard scale space axioms, discussed below, leads to the linear Gaussian scale-space, which is the most common type of scale space used in image processing and computer vision. == Scale space axioms for the linear scale-space representation == The linear scale space representation L ( x , y , t ) = ( T t f ) ( x , y ) = g ( x , y , t ) ∗ f ( x , y ) {\displaystyle L(x,y,t)=(T_{t}f)(x,y)=g(x,y,t)f(x,y)} of signal f ( x , y ) {\displaystyle f(x,y)} obtained by smoothing with the Gaussian kernel g ( x , y , t ) {\displaystyle g(x,y,t)} satisfies a number of properties 'scale-space axioms' that make it a special form of multi-scale representation: linearity T t ( a f + b h ) = a T t f + b T t h {\displaystyle T_{t}(af+bh)=aT_{t}f+bT_{t}h} where f {\displaystyle f} and h {\displaystyle h} are signals while a {\displaystyle a} and b {\displaystyle b} are constants, shift invariance T t S ( Δ x , Δ y ) f = S ( Δ x , Δ y ) T t f {\displaystyle T_{t}S_{(\Delta x,\Delta _{y})}f=S_{(\Delta x,\Delta _{y})}T_{t}f} where S ( Δ x , Δ y ) {\displaystyle S_{(\Delta x,\Delta _{y})}} denotes the shift (translation) operator ( S ( Δ x , Δ y ) f ) ( x , y ) = f ( x − Δ x , y − Δ y ) {\displaystyle (S_{(\Delta x,\Delta _{y})}f)(x,y)=f(x-\Delta x,y-\Delta y)} semi-group structure g ( x , y , t 1 ) ∗ g ( x , y , t 2 ) = g ( x , y , t 1 + t 2 ) {\displaystyle g(x,y,t_{1})g(x,y,t_{2})=g(x,y,t_{1}+t_{2})} with the associated cascade smoothing property L ( x , y , t 2 ) = g ( x , y , t 2 − t 1 ) ∗ L ( x , y , t 1 ) {\displaystyle L(x,y,t_{2})=g(x,y,t_{2}-t_{1})L(x,y,t_{1})} existence of an infinitesimal generator A {\displaystyle A} ∂ t L ( x , y , t ) = ( A L ) ( x , y , t ) {\displaystyle \partial _{t}L(x,y,t)=(AL)(x,y,t)} non-creation of local extrema (zero-crossings) in one dimension, non-enhancement of local extrema in any number of dimensions ∂ t L ( x , y , t ) ≤ 0 {\displaystyle \partial _{t}L(x,y,t)\leq 0} at spatial maxima and ∂ t L ( x , y , t ) ≥ 0 {\displaystyle \partial _{t}L(x,y,t)\geq 0} at spatial minima, rotational symmetry g ( x , y , t ) = h ( x 2 + y 2 , t ) {\displaystyle g(x,y,t)=h(x^{2}+y^{2},t)} for some function h {\displaystyle h} , scale invariance g ^ ( ω x , ω y , t ) = h ^ ( ω x φ ( t ) , ω x φ ( t ) ) {\displaystyle {\hat {g}}(\omega _{x},\omega _{y},t)={\hat {h}}({\frac {\omega _{x}}{\varphi (t)}},{\frac {\omega _{x}}{\varphi (t)}})} for some functions φ {\displaystyle \varphi } and h ^ {\displaystyle {\hat {h}}} where g ^ {\displaystyle {\hat {g}}} denotes the Fourier transform of g {\displaystyle g} , positivity g ( x , y , t ) ≥ 0 {\displaystyle g(x,y,t)\geq 0} , normalization ∫ x = − ∞ ∞ ∫ y = − ∞ ∞ g ( x , y , t ) d x d y = 1 {\displaystyle \int _{x=-\infty }^{\infty }\int _{y=-\infty }^{\infty }g(x,y,t)\,dx\,dy=1} . In fact, it can be shown that the Gaussian kernel is a unique choice given several different combinations of subsets of these scale-space axioms: most of the axioms (linearity, shift-invariance, semigroup) correspond to scaling being a semigroup of shift-invariant linear operator, which is satisfied by a number of families integral transforms, while "non-creation of local extrema" for one-dimensional signals or "non-enhancement of local extrema" for higher-dimensional signals are the crucial axioms which relate scale-spaces to smoothing (formally, parabolic partial differential equations), and hence select for the Gaussian. The Gaussian kernel is also separable in Cartesian coordinates, i.e. g ( x , y , t ) = g ( x , t ) g ( y , t ) {\displaystyle g(x,y,t)=g(x,t)\,g(y,t)} . Separability is, however, not counted as a scale-space axiom, since it is a coordinate dependent property related to issues of implementation. In addition, the requirement of separability in combination with rotational symmetry per se fixates the smoothing kernel to be a Gaussian. There exists a generalization of the Gaussian scale-space theory to more general affine and spatio-temporal scale-spaces. In addition to variabilities over scale, which original scale-space theory was designed to handle, this generalized scale-space theory also comprises other types of variabilities, including image deformations caused by viewing variations, approximated by local affine transformations, and relative motions between objects in the world and the observer, approximated by local Galilean transformations. In this theory, rotational symmetry is not imposed as a necessary scale-space axiom and is instead replaced by requirements of affine and/or Galilean covariance. The generalized scale-space theory leads to predictions about receptive field profiles in good qualitative agreement with receptive field profiles measured by cell recordings in biological vision. In the computer vision, image processing and signal processing literature there are many other multi-scale approaches, using wavelets and a variety of other kernels, that do not exploit or require the same requirements as scale space descriptions do; please see the article on related multi-scale approaches. There has also been work on discrete scale-space concepts that carry the scale-space properties over to the discrete domain; see the article on scale space implementation for examples and references.

    Read more →
  • Globetrooper

    Globetrooper

    Globetrooper is a free travel app known for assisting travelers in finding partners for group trips and world adventures. Globetrooper offers a free social travel platform that helps people find travel partners. == History == Globetrooper was developed and released in 2010 by a couple; Todd Sullivan and Lauren McLeod who are two travel-minded individuals that wanted to make it easier for travelers to plan a journey and see the world. With their backgrounds in business, software & design, and a love for travel, both left the corporate world and launched Globetrooper on Lauren’s birthday 28 March 2010. Globetrooper was first launched as an information portal with a view to making it more social, but after some months, the content quickly grew and changed to the ‘travel partner’ concept.

    Read more →
  • XLNet

    XLNet

    The XLNet was an autoregressive Transformer designed as an improvement over BERT, with 340M parameters and trained on 33 billion words. It was released on 19 June 2019, under the Apache 2.0 license. It achieved state-of-the-art results on a variety of natural language processing tasks, including language modeling, question answering, and natural language inference. == Architecture == The main idea of XLNet is to model language autoregressively like the GPT models, but allow for all possible permutations of a sentence. Concretely, consider the following sentence:My dog is cute.In standard autoregressive language modeling, the model would be tasked with predicting the probability of each word, conditioned on the previous words as its context: We factorize the joint probability of a sequence of words x 1 , … , x T {\displaystyle x_{1},\ldots ,x_{T}} using the chain rule: Pr ( x 1 , … , x T ) = Pr ( x 1 ) Pr ( x 2 | x 1 ) Pr ( x 3 | x 1 , x 2 ) … Pr ( x T | x 1 , … , x T − 1 ) . {\displaystyle \Pr(x_{1},\ldots ,x_{T})=\Pr(x_{1})\Pr(x_{2}|x_{1})\Pr(x_{3}|x_{1},x_{2})\ldots \Pr(x_{T}|x_{1},\ldots ,x_{T-1}).} For example, the sentence "My dog is cute" is factorized as: Pr ( My , dog , is , cute ) = Pr ( My ) Pr ( dog | My ) Pr ( is | My , dog ) Pr ( cute | My , dog , is ) . {\displaystyle \Pr({\text{My}},{\text{dog}},{\text{is}},{\text{cute}})=\Pr({\text{My}})\Pr({\text{dog}}|{\text{My}})\Pr({\text{is}}|{\text{My}},{\text{dog}})\Pr({\text{cute}}|{\text{My}},{\text{dog}},{\text{is}}).} Schematically, we can write it as → My → My dog → My dog is → My dog is cute . {\displaystyle {\texttt {}}{\texttt {}}{\texttt {}}{\texttt {}}\to {\text{My }}{\texttt {}}{\texttt {}}{\texttt {}}\to {\text{My dog }}{\texttt {}}{\texttt {}}\to {\text{My dog is }}{\texttt {}}\to {\text{My dog is cute}}.} However, for XLNet, the model is required to predict the words in a randomly generated order. Suppose we have sampled a randomly generated order 3241, then schematically, the model is required to perform the following prediction task: is dog is dog is cute → My dog is cute {\displaystyle {\texttt {}}{\texttt {}}{\texttt {}}{\texttt {}}\to {\texttt {}}{\texttt {}}{\text{is }}{\texttt {}}\to {\texttt {}}{\text{dog is }}{\texttt {}}\to {\texttt {}}{\text{dog is cute}}\to {\text{My dog is cute}}} By considering all permutations, XLNet is able to capture longer-range dependencies and better model the bidirectional context of words. === Two-Stream Self-Attention === To implement permutation language modeling, XLNet uses a two-stream self-attention mechanism. The two streams are: Content stream: This stream encodes the content of each word, as in standard causally masked self-attention. Query stream: This stream encodes the content of each word in the context of what has gone before. In more detail, it is a masked cross-attention mechanism, where the queries are from the query stream, and the key-value pairs are from the content stream. The content stream uses the causal mask M causal = [ 0 − ∞ − ∞ … − ∞ 0 0 − ∞ … − ∞ 0 0 0 … − ∞ ⋮ ⋮ ⋮ ⋱ ⋮ 0 0 0 … 0 ] {\displaystyle M_{\text{causal}}={\begin{bmatrix}0&-\infty &-\infty &\dots &-\infty \\0&0&-\infty &\dots &-\infty \\0&0&0&\dots &-\infty \\\vdots &\vdots &\vdots &\ddots &\vdots \\0&0&0&\dots &0\end{bmatrix}}} permuted by a random permutation matrix to P M causal P − 1 {\displaystyle PM_{\text{causal}}P^{-1}} . The query stream uses the cross-attention mask P ( M causal − ∞ I ) P − 1 {\displaystyle P(M_{\text{causal}}-\infty I)P^{-1}} , where the diagonal is subtracted away specifically to avoid the model "cheating" by looking at the content stream for what the current masked token is. Like the causal masking for GPT models, this two-stream masked architecture allows the model to train on all tokens in one forward pass. == Training == Two models were released: XLNet-Large, cased: 110M parameters, 24-layer, 1024-hidden, 16-heads XLNet-Base, cased: 340M parameters, 12-layer, 768-hidden, 12-heads. It was trained on a dataset that amounted to 32.89 billion tokens after tokenization with SentencePiece. The dataset was composed of BooksCorpus, and English Wikipedia, Giga5, ClueWeb 2012-B, and Common Crawl. It was trained on 512 TPU v3 chips, for 5.5 days. At the end of training, it still under-fitted the data, meaning it could have achieved lower loss with more training. It took 0.5 million steps with an Adam optimizer, linear learning rate decay, and a batch size of 8192.

    Read more →
  • CLEVER score

    CLEVER score

    The CLEVER (Cross Lipschitz Extreme Value for nEtwork Robustness) score is a way of measuring the robustness of an artificial neural network towards adversarial attacks. It was developed by a team at the MIT-IBM Watson AI Lab in IBM Research and first presented at the 2018 International Conference on Learning Representations. It was mentioned and reviewed by Ian Goodfellow as well. It was adopted into an educational game Fool The Bank by Narendra Nath Joshi, Abhishek Bhandwaldar and Casey Dugan

    Read more →
  • Deep learning

    Deep learning

    In machine learning, deep learning (DL) focuses on utilizing multilayered neural networks to perform tasks such as classification, regression, and representation learning. The field takes inspiration from biological neuroscience and revolves around stacking artificial neurons into layers and "training" them to process data. The adjective "deep" refers to the use of multiple layers (ranging from three to several hundred or thousands) in the network. Methods used can be supervised, semi-supervised or unsupervised. Some common deep learning network architectures include fully connected networks, deep belief networks, recurrent neural networks, convolutional neural networks, generative adversarial networks, transformers, and neural radiance fields. These architectures have been applied to fields including computer vision, speech recognition, natural language processing, machine translation, bioinformatics, drug design, medical image analysis, climate science, material inspection and board game programs, where they have produced results comparable to and in some cases surpassing human expert performance. Early forms of neural networks were inspired by information processing and distributed communication nodes in biological systems, particularly the human brain. However, current neural networks do not intend to model the brain function of organisms, and are generally seen as low-quality models for that purpose. == Overview == Most modern deep learning models are based on multi-layered neural networks such as convolutional neural networks and transformers, although they can also include propositional formulas or latent variables organized layer-wise in deep generative models such as the nodes in deep belief networks and deep Boltzmann machines. Fundamentally, deep learning refers to a class of machine learning algorithms in which a hierarchy of layers is used to transform input data into a progressively more abstract and composite representation. For example, in an image recognition model, the raw input may be an image (represented as a tensor of pixels). The first representational layer may attempt to identify basic shapes such as lines and circles, the second layer may compose and encode arrangements of edges, the third layer may encode a nose and eyes, and the fourth layer may recognize that the image contains a face. Importantly, a deep learning process can learn which features to optimally place at which level on its own. Prior to deep learning, machine learning techniques often involved hand-crafted feature engineering to transform the data into a more suitable representation for a classification algorithm to operate on. In the deep learning approach, features are not hand-crafted and the model discovers useful feature representations from the data automatically. This does not eliminate the need for hand-tuning; for example, varying numbers of layers and layer sizes can provide different degrees of abstraction. The word "deep" in "deep learning" refers to the number of layers through which the data is transformed. More precisely, deep learning systems have a substantial credit assignment path (CAP) depth. The CAP is the chain of transformations from input to output. CAPs describe potentially causal connections between input and output. For a feedforward neural network, the depth of the CAPs is that of the network and is the number of hidden layers plus one (as the output layer is also parameterized). For recurrent neural networks, in which a signal may propagate through a layer more than once, the CAP depth is potentially unlimited. No universally agreed-upon threshold of depth divides shallow learning from deep learning, but most researchers agree that deep learning involves CAP depth higher than two. CAP of depth two has been shown to be a universal approximator in the sense that it can emulate any function. Beyond that, more layers do not add to the function approximator ability of the network. Deep models (CAP > two) are able to extract better features than shallow models and hence, extra layers help in learning the features effectively. Deep learning architectures can be constructed with a greedy layer-by-layer method. Deep learning helps to disentangle these abstractions and pick out which features improve performance. Deep learning algorithms can be applied to unsupervised learning tasks. This is an important benefit because unlabeled data is more abundant than labeled data. Examples of deep structures that can be trained in an unsupervised manner are deep belief networks. The term deep learning was introduced to the machine learning community by Rina Dechter in 1986, and to artificial neural networks by Igor Aizenberg and colleagues in 2000, in the context of Boolean threshold neurons. The etymology of the term is more complicated. == Interpretations == Deep neural networks are generally interpreted in terms of the universal approximation theorem or probabilistic inference. The classic universal approximation theorem concerns the capacity of feedforward neural networks with a single hidden layer of finite size to approximate continuous functions. In 1989, the first proof was published by George Cybenko for sigmoid activation functions and was generalised to feed-forward multi-layer architectures in 1991 by Kurt Hornik. Recent work also showed that universal approximation also holds for non-bounded activation functions such as Kunihiko Fukushima's rectified linear unit. The universal approximation theorem for deep neural networks concerns the capacity of networks with bounded width but the depth is allowed to grow. Lu et al. proved that if the width of a deep neural network with ReLU activation is strictly larger than the input dimension, then the network can approximate any Lebesgue integrable function; if the width is smaller or equal to the input dimension, then a deep neural network is not a universal approximator. The probabilistic interpretation derives from the field of machine learning. It features inference, as well as the optimization concepts of training and testing, related to fitting and generalization, respectively. More specifically, the probabilistic interpretation considers the activation nonlinearity as a cumulative distribution function. The probabilistic interpretation led to the introduction of dropout as regularizer in neural networks. The probabilistic interpretation was introduced by researchers including Hopfield, Widrow and Narendra and popularized in surveys such as the one by Bishop. == History == === Before 1980 === There are two types of artificial neural network (ANN): feedforward neural network (FNN) or multilayer perceptron (MLP) and recurrent neural networks (RNN). RNNs have cycles in their connectivity structure, whereas FNNs do not. In the 1920s, Wilhelm Lenz and Ernst Ising created the Ising model which is essentially a non-learning RNN architecture consisting of neuron-like threshold elements. In 1972, Shun'ichi Amari made this architecture adaptive. His learning RNN was republished by John Hopfield in 1982. Other early recurrent neural networks were published by Kaoru Nakano in 1971. Already in 1948, Alan Turing produced work on "Intelligent Machinery" that was not published in his lifetime, containing "ideas related to artificial evolution and learning RNNs". Frank Rosenblatt (1958) proposed the perceptron, an MLP with 3 layers: an input layer, a hidden layer with randomized weights that did not learn, and an output layer. He later published a 1962 book that also introduced variants and computer experiments, including a version with four-layer perceptrons "with adaptive preterminal networks" where the last two layers have learned weights (here he credits H. D. Block and B. W. Knight). The book cites an earlier network by R. D. Joseph (1960) "functionally equivalent to a variation of" this four-layer system (the book mentions Joseph over 30 times). Should Joseph therefore be considered the originator of proper adaptive multilayer perceptrons with learning hidden units? Unfortunately, the learning algorithm was not a functional one, and fell into oblivion. The first working deep learning algorithm was the Group method of data handling, a method to train arbitrarily deep neural networks, published by Alexey Ivakhnenko and Lapa in 1965. They regarded it as a form of polynomial regression, or a generalization of Rosenblatt's perceptron to handle more complex, nonlinear, and hierarchical relationships. A 1971 paper described a deep network with eight layers trained by this method, which is based on layer by layer training through regression analysis. Superfluous hidden units are pruned using a separate validation set. Since the activation functions of the nodes are Kolmogorov-Gabor polynomials, these were also the first deep networks with multiplicative units or "gates". The first deep learning multilayer perceptron trained by stochastic gradient descent was published in 1967 by Shun'ichi

    Read more →